MULTIVOXEL MR SPECTROSCOPY TOOL
FOR BRAIN CANCER DETECTION IN NEURONAVIGATION
Performance
Juan José Fuertes
1
, Valery Naranjo
1
, Pablo González
1
, Ángela Bernabeu
2
, Mariano Alcañiz
1,3
and
Javier Sanchez
4
1
Instituto Interuniversitario de Investigación en Bioingeniería y Tecnología Orientada al Ser Humano (I3BH),
Universitat Politècnica de València, I3BH/LabHuman, Camino de Vera s/n, 46022, Valencia, Spain
2
Inscanner S.L., Unidad de Resonancia Magnética, Alicante, Spain
3
Ciber, Fisiopatología de Obesidad y Nutrición, CB06/03 Instituto de Salud Carlos III, Madrid, Spain
4
Philips Healthcare España, Maria de Portugal 1, 28050, Madrid, Spain
Keywords: Magnetic Resonance Spectroscopy, Brain Cancer Detection, Multivoxel, Magnetic Resonance Imaging,
Neuronavigator.
Abstract: This work presents a simple and interactive spectroscopic tool to help clinicians for brain cancer detection.
Firstly, Magnetic Resonance Spectroscopy (MRS) and Magnetic Resonance Imaging (MRI) are registered
to perform brain analysis. After processing the spectroscopic signals with HLSVD method for water
suppression, zero-filling and phase-correction algorithms, and apodization functions to improve the signal-
to-noise ratio (SNR), the metabolite brain maps are generated in order to analyze brain composition. A 3D-
spatial distribution of the anatomical and spectroscopic images and how they are registered are presented to
facilitate surgery planning. The goal is to generate metabolite brain maps which can be merged with
anatomical images in the neuronavigator to provide the surgeon with the exact point where performing the
biopsy.
1 INTRODUCTION
Many spectroscopic tools have been developed to
solve the problem of quantifying signals in
1
H MRS
data in the last 20 years (Castillo et al., 1996;
Stoyanova et al., 1995). The most popular tools are
the jMRUI software package (Stefan et al., 2009),
the LCModel
TM
(Gruber et al., 2004) and the
AQSES software (Simonetti et al., 2006) among a
large series of software applications. All of them
incorporate the main algorithms to process the
spectroscopic signals in order to correct distortions
and imperfections, perform water suppression and
generate metabolite maps (Poullet et al., 2007).
They offer black box quantitation algorithms based
on singular value decomposition (SVD) such as
HSVD (De Beer and Van Ormondt, 1992), HLSVD
(Pijnappel et al., 1992) and HTLS (Van Huffel et al.,
1994), which are efficient to quantify signals with
good signal-to-noise ratio (SNR); in addition, they
are used to suppress the dominant signals, for
example the water coin. Moreover, these tools also
offer quantitation algorithms with available prior
knowledge such as AMARES (Vanhamme and Van
Huffel, 1997) or QUEST (Ratiney et al., 2005),
which let us impose conditions on the model-
function parameters. There are much more methods
for spectroscopic analysis, but in summary, the main
steps included in most of software tools (Poullet et
al., 2007) in order to quantify the signals and
generate the metabolite maps are:
1. Phase correction.
2. Dominant signals removal.
3. Signal enhancement with noise reduction.
4. Estimation of spectral parameters.
5. Quantitation with model functions or signal
basis sets.
There are many frameworks which already exist,
therefore there are no need to re-invent the wheel but
understand processes, appreciate limitations and
167
Fuertes J., Naranjo V., González P., Bernabeu Á., Alcañiz M. and Sanchez J..
MULTIVOXEL MR SPECTROSCOPY TOOL FOR BRAIN CANCER DETECTION IN NEURONAVIGATION - Performance.
DOI: 10.5220/0003770501670172
In Proceedings of the International Conference on Biomedical Electronics and Devices (BIODEVICES-2012), pages 167-172
ISBN: 978-989-8425-91-1
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
provide clinicians with the needs they have. The
software we present incorporates the main
algorithms to process the data, performs the
registration between spectroscopic data and MR
images, and generates the metabolite maps.
Additionally, it introduces a 3D view to merge the
3D anatomical information reconstructed from the
MR images with the metabolite information from the
spectroscopic signals in order to satisfy clinicians. It
combines the effect of molecular structure on the
magnetic field experienced by an atomic nucleus,
the Chemical Shift, with the effects of the magnetic
field gradients used in MRI. In addition, the MRS
has enormous potential because it allows direct
observation of the chemical basis of disease.
In this work we are focusing on multivoxel
spectroscopy analysis instead of single voxel
spectroscopy in order to study the features of
multiple voxels at the same time. This is significant
when the area to study is not well defined, such as
tumour monitoring or biopsy surgery, where the
spatial distribution has notable importance. It
provides the clinicians with information about the
physical constitution of each disease.
In short, the aim of this work is to provide the
surgeon with a 3D view about the area where the
patient has the tumour and to generate a DICOM
metabolite image which will be used in the
neuronavigator when the surgeon will perform the
biopsy. In this way, it will be able to detect clearly
the tumour area.
This paper is set up as follows: section 2 shows
the software structure, explaining briefly each of the
four blocks of the tool: multivoxel selection, signal
processing, metabolite brain maps and 3D image
reconstruction view. In section 3 the features of the
algorithms to process the signals are detailed and
section 4 shows the image registration to provide the
3D brain view; in section 5, an acceptability test of
the software is shown. Finally, a brief conclusion is
given in section 6 and the future work is introduced
in section 7.
2 TOOL WINDOW STRUCTURE
This section explains how the application for brain
cancer detection is organized. It allows the surgeon
to see “the inside” of the anatomical patient’s brain
together with the metabolite maps after the signals
are analyzed. The tool is divided into four blocks
(Figure 1). In the first block, anatomical images are
shown in order to select the voxels of the signals
which will be processed. In the second block, the
spectroscopic signals are processed in order to
generate the metabolite maps which are presented
with the MR images in the third block. Finally, the
fourth block shows a 3D-view of the anatomical
structure of the patient’s brain and the field of view
(FOV) of the spectroscopy imaging (also the VOI
that stands for the volume of interest excited inside
the FOV) which let the surgeon see the surgery
planning to perform the biopsy.
Figure 1: Block diagram of the software tool.
2.1 Anatomical Images and Multivoxel
Selection
The first step once the application is running is to
load the anatomical images of the patient, providing
the user with the axial, coronal and sagittal images
so as to select those slices which are going to be
analyzed with spectroscopy imaging. Next, the
spectroscopic data are loaded and we are able to
select those voxels which belong to the tumour area.
Figure 2: Main window of the tool.
Multi-voxel selection Signal processing
Metabolite maps 3D view
BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices
168
The spectroscopy imaging has a size of H*W*P
millimetres, where H stands for height, W stands for
width and P stands for depth. In this way, the size of
each voxel can be calculated. Figure 2 shows the
main window where we can see the anatomical
images with the spectroscopic voxels. The red box is
the field of view of the spectroscopy imaging and
the red box belongs to the area of those voxels
which are excited. Both boxes are registered with the
anatomical images as we will show in section 4.
2.2 Signal Processing
After the voxels of spectroscopic images have been
selected, the signals are processed to quantify and
generate the metabolite maps. In Figure 3 it is
possible to see the processing block interface, where
water suppression, signal to noise improvement and
quantifying algorithms are applied.
Figure 3: Signal processing window.
This block is divided into two windows: in the
first one, the original signals in time and frequency
domain are shown. In x-axis, milliseconds, hertz,
parts per millon (ppm) and points are the existing
options; real, imaginary and module options are
available in y-axis. In the second block, the result
after processing the signals is given, calculating the
metabolite amplitudes to create the colour maps.
There are much more tools such as zoom, inverse
ilustration, etc. and it is also possible to analyze each
signal independently clicking on it.
2.3 Metabolite Brain Maps
Once the original signals are processed and
quantified, the metabolite maps are shown as a blue-
red scale in order to emphasize the areas where there
is carcinogenic tissue. In Figure 4 the multivoxel
selection window and the result-window are shown
where the metabolite maps are registered with
anatomical images. There is also the option to
generate a medical report automatically in pdf
format with the patient information, the disease
features, the anatomical images, the metabolite
maps, and the information that the clinicians wish.
Figure 4: Creation of metabolite maps.
2.4 3D Anatomical View - Surgery
Planning
In order to know the planning of the MR
spectroscopy, the 3D model of patient’s head
(Figure 5) has been reconstructed and the
registration with the volume of spectroscopy
imaging (FOV and VOI) performed. In this way, the
surgeon knows exactly the area to study with
spectroscopic technique and where the tumour is
located.
Figure 5: 3D brain reconstruction with spectroscopic
volume.
In addition, once the signals are processed and
the colour maps generated, it is also possible to
recover the 3D colour maps to see the dangerous
regions together with anatomical information (see
Figure 6).
Figure 6: 3D metabolite map reconstruction.
MULTIVOXEL MR SPECTROSCOPY TOOL FOR BRAIN CANCER DETECTION IN NEURONAVIGATION -
Performance
169
If all metabolite maps corresponding to
anatomical images are generated, the 3D colour
reconstruction of the full spectroscopic imaging is
performed. In this way the surgeon knows as a
colour scale the aggressiveness of the tumour.
3 SIGNAL PROCESSING
METHODS
The purposes of signal processing methods
incorporated in this application are water removal,
baseline and phased correction, signal to noise
improvement and signal quantitation. With this aim,
many algorithms have been proposed since 1990 in
the time and frequency domain (Vanhamme et al.,
(2011), but nowadays only some of them are really
used in multivoxel spectroscopy imaging. Ideally,
the FID signal obtained with the MRS machine is
noiseless and results exactly from the addition of K
exponentially damped sinusoids which are
characterized by frequencies f
k
, amplitudes A
k
,
phases φ
k
, damping factors α
k
, length of the FID N, i
the square root of -1 and Δt the sampling interval, as:
X
=
A
e

e




n= 0,...,N-1. (1)
Before applying the water removal algorithm, it
is also possible to correct the phase of each signal in
order to know the sign of the peaks; moreover, it has
to take into account because phase is the main factor
that affects resolution. In order to remove the water
signal, the exponentially damped sinusoids whose
frequencies appear in the water region are selected
and subtracted from the original FID through the
Hankel Lanczos Singular Value Decomposition
(HLSVD-PRO) algorithm (Laudadio et al., 2002),
which estimates the whole set of model parameters
making full use of mathematical model functions.
Optimal values of number of sinusoids (K) and
Hankel size (N and M) must be determined;
therefore, values obtained by computer simulations
are established by default (K=25, N=1024 and
M=512), allowing the clinicians to change these
values for each signal (Cabanes et al., 2001). The
user can select the water region with two cursors as well.
After the water signal is suppressed, Lorentzian
and Gaussian functions to improve the signal to
noise ratio are applied. The exponentially function
implemented to improve de SNR is:

=e


∗∗
(2)
where factor is the apodization coefficient, sign is
the sign of the factor, and t is the time which is t
2
when the function is Gaussian.
At this time, the quantitation method is
performed in order to know the concentration of the
metabolites in each voxel. AMARES and HLSVD
are the algorithms which can be used in our
application. In Figure 7 the signal fitting with
HLSVD method is shown. As the concentration of
each metabolite is known, it is possible to determine
the relationship among them. In this way, the colour
maps of Naa/Cr, Cho/Cr, Cho/Naa, etc. are
calculated (N-acetylaspartate (Naa), creatine (Cr),
cholina (Cho)).
Figure 7: Signal quantitation with HLSVD method.
The new metabolite maps are saved in DICOM
format images to be used in the neuronavigator. In
this way, the red areas where the tumour is
aggressive will be highlight to let the surgeon
perform the biopsy accurately.
4 IMAGE REGISTRATION
The 3D anatomical view was introduced in section
2.4 to show the new functionality included in this
software tool. in this manner, the surgeon is able to
see the sagittal, coronal and axial views and the 3D
view of the spectroscopic planning (Figure 8).
Figure 8: Sagittal, axial, coronal and 3D view.
NAA
Cr Cho
BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices
170
The first step for image registration is to locate
both anatomical and spectroscopic images in the 3D
space with their own rotation matrix. At this point, it
is possible to know which anatomical images belong
to each spectroscopic slice, through the expression:
, ,
=c
x, y, z
+t∗M’ (3)
where the absolute position of the each
spectroscopic slice is calculated. e(x,y,z) is the
absolute position of the edges of the spectrum,
c(x,y,z) is the center of the FOV, t is the thickness of
each slice and M is the rotation matrix. Thus, the
surgeon knows exactly the anatomical areas which
have been analyzed with MR spectroscopy.
5 ACCEPTABILITY TEST
One of the most important features when we deal
with medical software is the clinicians’ acceptability
besides the required accuracy. In order to make
easier the use of this application, we have been
working in the design of a user-friendly interface,
asking doctors for ease of use (green), noticed
usability (blue), human-machine interaction (yellow)
and satisfaction level (orange). After creating the
first version of the software, the obtained results in a
10 point scale after asking 5 radiologists and
medical engineers are:
Table 1: Results of the acceptability test.
Clarity when using methods 8
Software learning (easy) 9.4
Methodology to process data 9
Direct methods route 10
Signal processing methods (number and
suitable)
8.4
Image processing methods (number and
suitable)
9
Generate a PDF report 10
Time-work relation 8.6
Human-machine interaction 10
User friendly interface 10
I would like to use this tool in my work 9
Graphical results 9.4
Easy 9
Intuitive 8.6
Useful 9
Handy 9
The table 1 shows the evaluation results of the
software in terms of acceptability: the user
satisfaction and the human interaction prove the
good performance of the tool. We have to improve
the noticed usability in order to make the tool more
suitable.
6 CONCLUSIONS
In this paper, a software tool to help surgeons detect
brain cancer through the use of magnetic resonance
spectroscopy has been presented.
Many applications and algorithms have been
proposed in the last years, but only some of them are
really used. The application presented in this paper
integrates whose algorithms that are daily used
providing the clinician with a user-friendly tool.
Furthermore, our goal is not to re-invent the
algorithms but understand processes and provide the
clinicians with the needs they have. This is the
reason why a novel 3D planning view which shows
the area analyzed with MRS is incorporated. It
consists of anatomical and spectroscopic images,
and the 3D reconstruction of patient’s brain.
Another significant feature of the tool is the
creation of metabolite brain maps. They must be
compatible with neuronavigator, so the surgeon
knows exactly the point where pointing with a
needle when it is performing the surgery. For this
reason the images are saved in DICOM format.
7 FUTURE WORK
Coming soon experiments are focused on making a
comparison with other software to verify the
functionality of the algorithms in order to generate
reliable metabolite maps. As far as new tool methods
are concerned, the fusion with functional images
such as Positron Emission Tomography (PET) is
being studied. The goal is to perform a 3D PET
reconstruction and register anatomical, metabolite
and functional images.
Another area of study is the creation of DICOM
images which are compatible with neuronavigator.
In this way, we are working on the enhancement of
the tumour areas to provide the surgeon with a
friendly interface to perform the biopsy accurately.
Once the DICOM images are generated, the surgeon
will perform the biopsy in the operation theatre,
taking samples of the emphasized brain tissue and
corroborating if the patient has or not tumour. PET
MULTIVOXEL MR SPECTROSCOPY TOOL FOR BRAIN CANCER DETECTION IN NEURONAVIGATION -
Performance
171
will be also fused with the rest of images in the
neuronavigator.
New quantification and water removal methods
based on non-linear filters (1D mathematical
morphology) are also being studied for its
application to MR spectroscopy.
ACKNOWLEDGEMENTS
This work has been supported by Centro para el
Desarrollo Tecnológico Industrial (CDTI) under the
project ONCOTIC (IDI-20101153), and partially by
projects Consolider-C (SEJ2006-14301/PSIC),
“CIBER of Physiopathology of Obesity and
Nutrition, an initiative of ISCIII” and Excellence
Research Program PROMETEO (Generalitat
Valenciana. Conselleria de Educación, 2008-157).
We would like to express our deep gratitude to the
Hospital Clínica Benidorm for its participation in
this project. The work of Juan José Fuertes has been
supported by a FPI grant from “Programa de Ayudas
de Investigación y Desarrollo (PAID)” of UPV.
REFERENCES
Cabanes, E., Confort-Gouny, S., Le Fur, Y., Simond, G.
and Cozzone, P.J. (2001). Optimization of Residual
Water Signal Removal by HLSVD on Simulated Short
Echo Time Proton MR Spectra of the Human Brain.
Journal of Magnetic Resonance. 150:116-125.
Castillo, M., Kwock, L., and Mukherji, S. (1996). Clinical
application of proton MR spectroscopy. American
Journal of Neuroradiology. 17:1-15.
De Beer, R. and Van Ormondt, D. (1992). Analysis of
NMR data using time domain fitting procedures. NMR
Basic Principles and Progress. 26:202-48.
Gruber, S., Stadibauer, A., Mlynarik, V., Ganslandt, O.,
and Moser, E. (2004). An LCModel-based automatic
software tool for the reconstruction of anatomically-
and pathologically-matched voxels using high-
resolution 3D-spectroscopic imaging: Applications in
tumour patients. In Proc. ISMRM, volume 11.
Laudadio, T., Mastronardi, N., Vanhamme, L., Van
Hecke, P. and Van Huffel, S. (2002). Improved
Lanczos algorithms for blackbox MRS data
quantitation. Journal of Magnetic Resonance.
157:292-297.
Pijnappel, W. W. F., Van den Boogaart, A., De Beer, R.,
and Van Ormondt, D. (1992). SVD-based
quantiffication of magnetic resonance signals. Journal
of Magnetic Resonance. 97(1):122-134.
Poullet, J., Sima, D. M., Van Huffel, S., and Van Hecke,
P. (2007). Frequency-selective quantitation of short-
echo time 1H magnetic resonance spectra. Journal of
Magnetic Resonance. 186(2):293-304,
Ratiney, H., Sdika, M., Coenradie, Y., Cavassila, S., Van
Ormondt, D., and Graveron-Demilly, D. (2005). Time
domain semi-parametric estimation based on a
metabolite basis set. NMR in Biomedicine. 18(1):1-13.
Simonetti, A. W., Poullet, J-B., Sima, D. M., De Neuter,
B., Vanhamme, L., Lemmerling, P., and Van Huffel S.
(2006). An open source short echo time MR
quantitation software solution: AQSES.
Stefan, D., Di Cesare, F., Andrasescu, A., Popa, E.,
Lazariev, A., Vescovo, E., Strbak, O., Williams, S.,
Starcuk, Z., Cabanas, M., Van Ormondt, D., and
Graveron-Demilly, D. (2009). Quantitation of
magnetic resonance spectroscopy signals: the jMRUI
software package. Measurement Science and
Technology, volume 20.
Stoyanova, R., Kuesel, A., and Brown, T. (1995)
Application of principal-component analysis for
spectral Quantitation. Journal of Magnetic Resonance.
115:265-269.
Van Huffel, S., Chen, H., Decanniere, C., and Van Hecke,
P. (1994) Algorithm for time-domain NMR data
fitting based on total least squares. Journal of
Magnetic Resonance. 110:228-237.
Vanhamme, L., and Van Huffel, S. (1997). Improved
method for accurate and efficient quantification of
MRS data with use of prior-knowledge. Journal of
Magnetic Resonance. 129(1):35-43.
Vanhamme, L., Sundin, T., Hecke, P. V., and Huffel, S. V.
(2001). MR Spectroscopic Quantitation: a Review of
Time-Domain Methods. NMR in Biomedicine. 14:233-
246.
BIODEVICES 2012 - International Conference on Biomedical Electronics and Devices
172